Metabolism of iodine-123-BMIPP in perfused rat hearts. J NucÃ-Med 1995;36:1043-1050. 16. Tamaki N. Kawamoto M, Yonekura Y. et al. Decreased uptake of ...
12. Takeishi Y. Sukekawa H, Saito H, et al. Clinical significance of decreased myocardial uptake of I:!M-BMIPP in patients with stable effort angina pectoris. NucÃ-Med Commun 1995; 16:1002-1008. 13. Takeishi Y. Sukekawa H, Fujiwara S, Ikeno E. Sasaki Y, Tomoike H. Reverse redistribution of WmTc-sestamibi following direct PTCA in acute myocardial infarc tion. JNucÃ- Med 1996;37:I289-1294. 14. Fujiwara S, Takeishi Y. Atsumi H, Takahashi K, Tomoike H. Fatty acid metabolic imaging for the diagnosis of coronary artery disease. J NucÃ-Med 1997; 37:175-180. 15. Yamamichi Y. Kusuoka H, Morishita K. et al. Metabolism of iodine-123-BMIPP in perfused rat hearts. J NucÃ-Med 1995;36:1043-1050.
16. Tamaki N. Kawamoto M, Yonekura Y. et al. Decreased uptake of 1-123 BMIPP as a sign of enhanced glucose utilization assessed by FDG-PET [Abstract], J NucÃ-Med 1991:32:1034. 17. GrÃ©goireJ. Theroux P. Detection and assessment of unstable angina using myocardial perfusion imaging: comparison between technetium-99m sestamibi SPECT and 12-lead electrocardiogram. Am J Cardio! 1990;66:42E-46E. 18. Bilodeau L, Theroux P, GrÃ©goireJ. Gagnon D. Arsenault A. Technetium-99m sestamibi tomography in patients with spontaneous chest pain: correlations with clinical, electrocardiographic and angiographie findings. J Am Coll Cardiol 1991:18:1684-1691.
Image Enhancement of Severely Hypoperfused Myocardia for Computation of Tomographie Ejection Fraction Kenneth Nichols. E. Gordon DePuey, Alan Rozanski, HÃ©lÃ¨ne Salensky and Marvin I. Friedman Department of Radiology and Division of Cardiology, St. Luke's-Roosevelt Hospital and Columbia University College of Physicians and Surgeons, New York, New York Ejection fractions computed from 99nTc-sestamibi myocardial per fusion gated tomograms have demonstrated a high degree of accuracy and reproducibility. Although automated algorithms ap pear to provide reasonable endocardial outlines for patients over a broad spectrum of cardiac diseases, in cases of severe hypoperfusion, it is necessary to manually adjust contrast and brightness to judge whether borders are correct or must be altered. Methods: Midventricular horizontal and vertical long axis gated tomograms were generated for 116 studies chosen on the basis of extensive, severe myocardial perfusion defects. Automated software trans formed cinematic tomograms into images demonstrating uniform appearance of the myocardium throughout the cardiac cycle. Trans formed images were introduced to edge detection algorithms for subsequent calculation of ventricular volumes and ejection frac tions. Results: Linear regression analysis demonstrated excellent intraobserver reproducibility for ejection fractions (r = 0.95) and volumes (r = 0.98). There was also good agreement of ejection fractions (r = 0.86) and volumes (r = 0.94) with values derived from an expert's manual drawings. In a subgroup of 22 patients, auto mated ejection fractions from transformed images demonstrated better agreement with independent first-pass values (r = 0.90) than did manual measurements derived from original data (r = 0.85). Conclusion: Image enhancement algorithms succeeded in provid ing accurate, reproducible gated SPECT ejection fractions in the most difficult class of patients exhibiting severe hypoperfusion. Key Words: gated SPECT; image enhancement; ejection fraction J NucÃ-Med 1997; 38:1411-1417
Ahe importance of left ventricular (LV) ejection fraction (EF) measurements in the clinical management of cardiac patients is well established (1,2). EF calculations using 99mTc-sestamibi myocardial perfusion gated tomograms have been reported using different approaches, including simple manual techniques (3,4), relaxation labeling combined with computed wall thick ening (5,6) and Gaussian midmyocardial detection (7). Excel lent reproducibility of gated SPECT EF has been reported (8), and these measurements correlate well with those obtained by gated blood pool equilibrium studies (9, JO), first-pass radionuclide angiography (7,9,11,12), echocardiography (lÃ¬),magnetic resonance imaging (5,14) and x-ray contrast ventriculography Received Sep. 12, 1996; revision accepted Feb. 3, 1997. For correspondence or reprints contact: Kenneth Nichols, PhD, Division of Cardiol ogy, St. Luke's-Rcosevelt Hospital, Amsterdam Avenue at 114th St., New York, NY 10025.
FIGURE 1. Automatic endocardial borders for a case of severe hypoperfu sion of the apico-lateral territories.
(4,11). Additionally, gated SPECT EF processing has been automated successfully (7,9), as has selection of LV limits and axes needed to provide input to gated SPECT EF software (15,16). Despite these validations and regardless of which computa tional method is used, proper interpretation of myocardial perfusion tomograms remains problematic in cases of severe myocardial hypoperfusion (Fig. 1). Endocardial borders shown in Figure 1 were generated automatically (9), and although these are plausible, it is difficult to ascertain whether these truly represent the actual myocardium because the apico-lateral wall is nearly invisible in horizontal long axis (HLA) images and because apical counts are greatly reduced in vertical long axis (VLA) views. In addition, greatly reduced counts can defeat algorithms, which are based on gated SPECT count densities, designed to detect heart walls. It is important to review images in recognizing and compen sating for artifacts in interpreting studies (17,18). Yet, some gated SPECT EF methods are largely divorced from a visual appreciation of original myocardial perfusion images, despite displaying final results in four dimensions (5,7). Some auto-
IMAGEENHANCEMENT FORCOMPUTATION OF EF â€¢ Nichols et al.
"defect extent" of 167 Â±58 and "defect severity" of 830 Â±394 (20). Defect count density was 7% Â±3% of normal maximum myocardial counts. First-Pass Imaging
Among patients studied were a subgroup of 22 (aged 60 Â±10 yr; 50% men) who also had gated first-pass (GFP) Anger camera data acquisitions of the 814 MBq of 99mTc-sestamibi bolus used for the resting portion of a 2-day sestamibi perfusion protocol (21). Their perfusion scans exhibited quantitative perfusion polar map defect extent of 152 Â±64 and defect severity of 719 Â±365 (20). Defect count density was 7% Â±3% of normal maximum myocardial counts. Quality assurance criteria were applied such that GFP studies were discarded if bolus transit time exceeded 2.0 sec or if lung transit time exceeded 10 heart beats. All GFP data were analyzed by software previously validated against both equilibrium gated blood pool and multicrystal first-pass gamma camera correl ative studies (22). Gated SPECT Imaging
FIGURE 2. Stress and rest perfusion polar maps corresponding to the myocardial perfusion images of Figure 1.
mated techniques show few intermediate steps during execution (7,76), whereas others rely on Fourier fits to time series of perfusion polar maps, which observers do not see as part of data review (5,6). Only one gated SPECT EF method altered the display of original gated tomograms through image inversion (11,19) but more as a preprocessing step to enhance edges than as an aid in visual interpretation. Visual verification of success of endocardial border identifi cation is always important, and in cases of severe hypoperfusion, observers must adjust contrast and brightness of computer displays and must do so manually, to decide whether they are satisfied with results and to aid in effecting edge changes. Therefore, we sought to simultaneously address issues of improving algorithmic endocardial border detection and of visual verification by automatically altering regional contrast of gated myocardial perfusion tomograms, our main goal being to augment quality assurance of automated EF calculations in cases of severe hypoperfusion. The approach that was chosen was to produce transformed images that optimally depict myocardial walls throughout the cardiac cycle. In the process, two types of "mapped" cinematic images were generated of use to observers: "brightening-mapping," for aiding visual appreci ation of myocardial wall thickening, and "motion-mapping," to facilitate discernment of regional wall translation associated with ventricular contraction. MATERIALS
Technetium-99m stress myocardial perfusion data from January 1, 1992, to December 31,1995, were reviewed retrospectively, and from this data, 123 patient studies were selected at random solely on the basis of exhibiting severe myocardial hypoperfusion, de fined by stress quantitative perfusion polar maps with defect counts below 15% of normal maximum counts, along with defect extent >25% of polar map area (Fig. 2). Seven of these studies were subsequently discarded when it was found that splanchnic activity greatly exceeded severely hypoperfused LV inferior wall myocar dial activity. No other criteria were imposed to select or exclude patient data. The 116 remaining patients (aged 66 Â±13 yr; 58% men) were characterized by quantitative perfusion polar map 1412
Technetium-99m-sestamibi injections were performed during peak exercise of a Bruce protocol with a treadmill using 1.11 GBq for a 1-day protocol or 814 MBq for a separate-day protocol (21). Tomograms (64 X 64) with a pixel size of 6.4 mm were acquired with high-resolution collimation for 20 sec at 64 projections over 180Â°with a biplane camera so that acquisitions were performed in 12 min. Tomograms were acquired of the stress perfusion distri bution with patients at rest, greater than 30 min after stress, at 8 frames per R-R interval with a 100% beat acceptance window. Standard clinical data processing parameters were used (20): Butterworth (0.40,10) prefilters for gated tomograms, followed by quantitative ramp x-filtering (23), interslice spatial averaging and time-filtering among the eight gated frames. Images were reori ented into VLA, HLA and short axis (SA) sections using manual choices by an observer of anterior, inferior, septal and lateral limits and approximate LV symmetry axes. Commercially available software (20) was used to produce stress and rest perfusion polar maps (Fig. 2) from SA images and to generate cinematic midventricular VLA and HLA images. These cinematic tomograms were input to previously described software that automatically generated endocardial borders on end-diastolic (ED) and end-systolic (ES) frames (9). An observer had the option of altering automated LV centers, ED and ES frame identification and endocardial borders. During automatic endocardial border generation (Fig. 1), estimated points of intersection of the valve plane with the myocardium from ED and ES images were output to a computer file for later use by image transformation algorithms described below. Subsequent transformations were performed from the valve plane forward to the apex, but for the valve plane to the outflow tract, transforma tions were modulated linearly to leave outflow tract counts unal tered and dark. Image Transformations
In the simple regional contrast transformation, referred to here as brightening-mapping, a radial search was performed in 2Â°incre ments at every angle, 0, of the centered ED VLA frame to find the local maximum count, Dmax(0) (Fig. 3). Then the factor, f(0), needed to boost Dmax(0)up to the maximum myocardial count of the entire image, Mmax, was computed as f(0) = Mmax/Dmax(0). Finally, the counts, c[r,0], in each radial pixel at fixed angle 0 were multiplied by that factor f(0) along the entire radius, r, of the input image to form the output transformed ED image. This mapping rendered the ED midmyocardium uniformly bright. The same set of ED angular factors, f(0), were then applied to all other frames to reveal relative percentage of regional brightening (Fig. 3). This was postulated to reflect myocardial thickening by way of well-known
THEJOURNAL OFNUCLEAR MEDICINE â€¢ Vol. 38 â€¢ No. 9 â€¢ September 1997
FIGURE 3. A schematic of the image transformations. The top left depicts the set of radial pixels through the apical VLA image at angle 0 = 0Â°,with a maximum count Dmax(e=0Â°)of 115 counts at the middle of the myocardium. The radial count profile along that ray is shown (dashed line) in the graph at the top right. This input curve was multiplied by the factor necessary to bring the maximum of the output curve (solid line) up to 255 counts. The bottom left depicts the set of radial pixels through the infero-apical 0 = 45Â°angle, with a maximum myocardial count Dmax(0=45Â°)of 202 counts. The radial count profile along the ray at 8 = 45Â°is shown (dashed line) in the graph at the bottom right, which was multiplied by a different factor needed to produce an output curve (solid line) at the same new 255 count maximum level. This process was performed at every angle in 2Â°incre ments. The input curves shown here are taken from the VLA images of Figure 1.
partial volume effects (24). Other investigators have computed polar myocardial thickening maps analogous to perfusion polar maps (6,19,25), but the implications of this approach had not previously been rendered visualizable. The midmyocardial ED HLA tomogram was then used to find a new set of factors, f(0), which were applied to all midmyocardial HLA cinematic frames. The motion-mapping algorithm differed from that of brightening-mapping only in that angular mapping factors f(0) were determined from and applied to each cinematic frame separately. Because the maximum radial count at every angle of every frame was used as the basis of normalizing each ray of each cinematic image, this transformation provided uniform appearance of the myocardium for every frame throughout the cardiac cycle, at the expense of any impression of brightening. Effects of brightening-mapping and motion-mapping are illus trated in Figure 4 for the same patient data input. In the example
RGURE 4. The results of the radial image transformations. The top row shows the original four cinematic midventricular VLA tomograms from ED on the left to ES on the right. The middle row consists of brightening-mapped output Â¡mages,and the bottom row displays corresponding motion-mapped images. The example shown is for the same patient of Figures 1-3.
shown, only the infero-apical territory brightened and only by a small amount. The corresponding ED brightening-mapped image demonstrated a uniform appearance of the myocardium, as ex pected, but at ES areas at lowest count density intensified in discrete, disconnected "globules." Anomalous count density "gaps" have also been observed in some brightening-mappingtransformed images. These globules and gaps were merely noise artifacts, which would be distracting if the goal were to identify either midmyocardial or endocardia! points. However, there were no noise artifacts for motion-mapped images at ES and, as expected, no systolic brightening (Fig. 4). Thus, for purposes of tracking endocardium, either visually or algorithmically, motionmapping provided more consistent and more easily interpretable input images than either original data or brightening-mapped images. For this reason, motion-mapped images and not brighten-
FIGURE 5. The use of HLA and VLA motion-mapped images for endocardial border generation is illustrated for a patient exhibiting severe hypoperfusion of the infero-apical territories. The top row shows the original images at ED and ES, the middle row shows the motion-mapped Â¡magesand the bottom row shows the automatically determined endocardial borders, as generated using the motion-mapped images as input.
IMAGEENHANCEMENT FORCOMPUTATION OF EF â€¢ Nichols et al.
TABLE 1 Intraobserver Reproducibility of Motion-Mapped Parameters: Linear Regression Analysis Parameter
TABLE 3 Automated Motion-Mapped Compared with Manual Unmapped Measurements: Linear Regression Results Parameter
EFEDVESVAEFAEDVAESV1161161161161161160.870.940.940.210.030.034.0%6.6ml-0. EFEDVESVAEFAEDVAESV1161161161161161160.950.980.980.070.010.030.9%5.0ml2.9ml-1.1%2.3ml0.8ml0.970.960.970.02-0.02-0.014.5%13.2ml13.0ml4.5%13.4ml13.1 ml20.5 ml-1.7%-1.9ml-6.0 ml7.0%23.1 ml20.9 ml0.960.930.930.11-0.01-0.017.2%22.6 ml
ing-mapped images were used for automated endocardial edge detection as shown in Figure 5. It was not necessary to alter existing edge-tracking software (9) to generate automatically the endocardial borders of Figure 5. Rather, it was sufficient to normalize input images to the appro priate maximum count level because previously written algorithms assumed input data would always have the same normalized maximum counts. The final VLA and HLA endocardial border dimensions were then corrected for the camera's line spread function and combined to compute ventricular volumes and EF values with the Simpson's rule formula (3). Manual Computations
In the absence of algorithms to enhance visual appearance of the myocardium, EF values computed from severely hypoperfused myocardial perfusion gated SPECT studies would have required careful review by an observer who was experienced in analysis of such data. Therefore, an observer who was already well versed in manual methods (3) reviewed VLA and HLA ED and ES images along with the cinematic playback of these data using simultaneous monochrome and color displays, while manually drawing endocar dial borders. The observer was free to alter contrast and brightness of all displays to derive as much helpful information as possible as to likely locations of the endocardium. These edges where then used for independent computation of LV volumes and EF values, as previously described (3,9).
dently on different occasions, with Pearson product-moment correlation coefficients, r, of 0.95 and with a slope of 0.98. Likewise, ED volume (EDV) and ES volume (ESV) reproducibility was very high (r = 0.98). Regression analysis of Bland-Altman plots revealed no significant trends, with abso lute values of slopes of